Building Data Science Teams

Last year I was invited onto the Evolution Exchange Podcast to discuss my thoughts on how to build effective data science teams to build AI. I was joined by 3 other data science leaders: Jon Shen, Head of Solutions Engineering At Arturo; Kale Miller, Founder of Prometheus Al; and Andrew Wilson, CEO of Advent Atum. We discussed topics ranging from optimal team composition, to organisational culture to recruitment and hiring. My learnings from the discussion highlight the importance of fostering an innovative culture, getting runs on the board early and to build out your team with generalists prior to specialising.

1. AI lives and dies with culture

I am not in the people department, nor could I be, but I believe that culture is the key to success with building AI. The (sad) truth is that the success of a model or an AI application relies on adoption, not on the quality of your model.

A business is made up of people with various backgrounds; and the majority do not have a stats degree like myself. So it is our job to demonstrate how AI can improve their lives, and make them more successful in there job. The culture you need to foster is one that can embrace change. AI changes businesses, and requires people to learn, fast.

It is also our job to demonstrate AI’s limitations. AI is believed to be all powerful but really it’s just a bunch of 1s and 0s. AI sounds and feels human, but it’s an imposter. When the business uncovers this, it is very easy to look outward and believe the AI is greener on the other side. In short, be honest, be visible and be vocal.

2. Start Small, Fast, and Simple

Don Bradman only hit 6x sixes in his entire career – paraphrasing, he said you can’t get caught if you hit the ball on the ground. Bradman amassed huge test scores by starting small – hitting singles. Data science can be thought of in the same way – it’s important to get your eye in early with some quick wins before you start to take on ambitious projects.

In the podcast, Jon, a former actuary, discussed the concept of a portfolio of projects. Each project had a probability of success ranging from almost certain, to a moonshot. He highlighted that moonshots are great, they have the ability to propel a business into a new paradigm; but they were also likely to fail. It is important to balance your portfolio – taking on quick wins while having 1 or 2 moonshot projects.

I highlighted that starting off, its best to use more “common” industry standard models to get concepts off the ground quickly. You can always improve on your models later, but you can’t get time back, and practical feedback is so important. You may also surprise yourself, with how well a linear model, or PCA algorithm works in production.

3. You don’t need experts (yet)

The final part of the discussion was how to compose your teams. Data Science teams come in many shapes and sizes, ranging levels of seniority and expertise. What was optimal was ultimately up to the business – however there were some common trends.

Small agile teams typically work best. Large monolithic data science teams often morph into inefficient BI teams. Generalists work best with small businesses. Specialising early can lead to missed opportunities; however as a business grows, and your tech stack hardens, bringing specialists on board can bring AI from good to great.

Finally, innovation can come from anywhere – from an entry level data scientist to a veteran professional. Striking a balance between fresh ideas and experience is important in driving an effective data science team. If innovation is your goal, remember you can’t have experience in something that has never been done before. Coaching staff to understand what’s important to consider rather than ‘how’ to do there job will foster a high performing and engaged team.

Link to full podcast: https://evolutionjobs.com/exchange/evo-au-159-how-to-build-teams-to-build-ai/